Jev, Type Safe’s first System One Model, promises calibrated probabilities for classification tasks without training data. The article argues that while Jev can produce useful scores, its calibration claim is flawed. Calibration depends on both the model and the data distribution; a model calibrated on one dataset may be miscalibrated on another.

The author highlights a coin‑flip example where Jev predicts a 0.92 probability of heads, showing severe miscalibration. The author recommends treating Jev’s outputs as ranking scores rather than reliable probabilities and suggests cheap recalibration using a few hundred labeled examples if precise probabilities are needed. The piece emphasizes that Jev’s value lies in its ability to classify without data, but users must verify calibration on their own data before relying on numeric thresholds or cost calculations.